Hidden unit 2

1
Artificial neural networks as analytic tools in
an event-related potential study of face
memory Reiko Graham Michael R.W. Dawson The
Biological Computation Project, University of
Alberta, Edmonton, AB, Canada
Although further research is needed in order to
establish a framework for future analyses,
results provide support for the utility of ANNs
for ERP analysis and classification.
ANOVA was unable to distinguish between the two
types of ERPs however, an ANN was. Network
interpretation revealed that classification was
achieved through coarse coding in the hidden
units. Differences between input time-points that
varied according to ERP type were also
discovered.
We compared the abilities of artificial neural
networks (ANNs) and ANOVA to classify early
latency event-related potentials (ERPs) that
were recorded from the right temporal area
elicited by recognized and novel faces.
Hidden unit 2
5. Discussion Early latency information in ERPs
from unimodal face processing areas can be used
to differentiate between hits and CRs, but this
information is in the form of higher-order,
non-linear voltage/time relationships.   The ANN
can discriminate between ERPs through a form of
coarse coding in the hidden units, which may be
analogous to encoding in the temporal cortex
where faces are represented by patterns of
activity in a small number of broadly tuned
neurons (Young, 1995).   Examination of the
correlations of inputs and net input to a hidden
unit revealed differential relationships
dependent upon the recognition status of an
face.   Much remains to be revealed about how
ANNs classify ERPs and a framework for future
analyses remains to be established. Nevertheless,
preliminary results are encouraging and provide
support for unimodal memory effects.
1. Rationale Memory can be conceptualized as
the result of processing in unimodal and
transmodal cortical association areas (Mesulam,
1998). Unimodal areas are modality specific and
receive projections from primary sensory cortex.
In humans, unimodal visual areas include the
fusiform, inferior, and middle temporal areas.
Transmodal areas receive inputs from more than
one modality and include the prefrontal and
posterior parietal cortices. Neuroimaging (e.g.
Kanwisher et al., 1998) and intracranial (e.g.
McCarthy, 1997) have identified the fusiform area
of the temporal lobe as a unimodal area that is
important in face processing. Electrophysiological
ly, perceptual activity in this area is
manifested in the N200 which is maximal over
right temporal areas (Bentin et al., 1996). An
important issue is whether activity in this area
is also correlated with memory. ERP evidence with
humans is mixed. Some studies have reported early
latency face memory effects over temporal areas
(e.g. Seeck, et al., 1997), while some have not
(e.g. Graham Cabeza, 2001 Muente et al.,
1997).   One possibility for this
inconsistency is that memory-related voltage
changes are represented in ERPs but linear
methods of analysis are unable to reliably detect
them. Objectives Can ANNs detect
differences between early latency ERPs recorded
over temporal sites that were elicited by old and
new faces?   If an ANN can differentiate between
ERPs, how is it doing it? What features of the
data appear to be important?  
3. Results RM-ANOVA did not detect any
reliable differences between ERPs to hits and
CRs during the first 500ms. The ANN was able to
differentiate between the two types.
Examination of hidden unit activity revealed that
units had relatively non-differential activity to
hits and CRs. However, when the three units were
examined together, it was possible to see how the
discrimination was achieved.   Given that
hidden unit activity is a function of net input,
we correlated inputs to a hidden unit with its
net input, enabling us to determine which
time-points had relationships with net input and
hence, which may have influenced hidden unit
activity. Correlations revealed
relationships which differed depending on ERP
type. Step-wise regression confirmed that subsets
of inputs accounted for a significant amount of
variance in hidden unit activity.  
2. Method ERPs were obtained for remembered
faces (hits) and new faces (correct rejections or
CRs) from 42 subjects during a face recognition
task. To examine unimodal effects, voltages were
taken from the right temporal site (T8). Early
ERPs were isolated by taking the first 500ms of
the recording epoch. Time-points were averaged
into 20ms epochs.  An RM-ANOVA was
conducted which included the 25 epochs as
predictors and trial type and epoch as
within-subjects variables.   An ANN was trained
using the 25 epochs as inputs. We employed a
hybrid ANN which used integration devices as
hidden units and a value unit as the output unit
(integration devices transform the data with a
sigmoid function, value units, with a Gaussian
function). The ANN had 25 input units, 3 hidden
units and 1 output unit.
Methods Subjects 42 right-handed healthy males
and females. Materials 240 black and white
photographs of unfamiliar male and female faces.
Half of the faces were presented during study,
and the remainder were used as distractor faces
during recognition. Procedure Each trial began
with the presentation of a fixation rectangle
which was replaced by a face for 400 msec. After
1600 msec, subjects were presented with a
response selection screen. Trials were separated
by one second. During test, subjects indicated
which faces were old and which were new ERP
methods ERPs were recorded from 30 Ag/AgCl
electrodes. EEG was sampled for an epoch of 1700
msec, starting 100 msec prior to the onset of a
face. References Bentin, S. et al. (1996).
Jnl of Cog Neurosci, 8(6), 551-565. Graham, R.
Cabeza, R. (2001). Neuroreport, 12(2),
245-248. Kanwisher, N. et al. (1998). Cognition,
68, B1-B11. McCarthy, G.et al. (1997). Jnl of
Cog Neurosci, 9(5), 605-610. Mesulam, M.-M.
(1998). Brain, 121, 1013-1052. Muente, T.F. et
al. (1997). Neuroscience Research, 28,
223-233. Seeck, M. et al.(1997). Neuroreport, 8,
2749-2754. Young, M.P. (1995). In M. Gazzaniga
(Ed.), The cognitive neurosciences (pp. 463-474)
. Cambridge, MA MIT Press.
R2 .91
R2 .84
Hidden unit 1
Hidden unit 2
R2 .74
Hidden unit 3